[D3D12] Don't use switch in shaders on Intel, log GPU name

src/xenia/gpu/d3d12/pipeline_cache.cc

@@ -41,7 +41,10 @@ PipelineCache::PipelineCache(D3D12CommandProcessor* command_processor,
     : command_processor_(command_processor),
       register_file_(register_file),
       edram_rov_used_(edram_rov_used) {
-  shader_translator_ = std::make_unique<DxbcShaderTranslator>(edram_rov_used_);
+  auto provider = command_processor_->GetD3D12Context()->GetD3D12Provider();
+
+  shader_translator_ = std::make_unique<DxbcShaderTranslator>(
+      provider->GetAdapterVendorID(), edram_rov_used_);
 
   if (edram_rov_used_) {
     depth_only_pixel_shader_ =

src/xenia/gpu/dxbc_shader_translator.cc

@@ -33,7 +33,9 @@ DEFINE_bool(dxbc_switch, true,
             "on may improve stability, though this heavily depends on the "
             "driver - on AMD, it's recommended to have this set to true, as "
             "Halo 3 appears to crash when if is used for flow control "
-            "(possibly the shader compiler tries to flatten them).");
+            "(possibly the shader compiler tries to flatten them). On Intel "
+            "HD Graphics, this is ignored because of a crash with the switch "
+            "instruction.");
 DEFINE_bool(dxbc_source_map, false,
             "Disassemble Xenos instructions as comments in the resulting DXBC "
             "for debugging.");
@@ -85,8 +87,9 @@ constexpr uint32_t
 constexpr uint32_t DxbcShaderTranslator::kCbufferIndexUnallocated;
 constexpr uint32_t DxbcShaderTranslator::kCfExecBoolConstantNone;
 
-DxbcShaderTranslator::DxbcShaderTranslator(bool edram_rov_used)
+DxbcShaderTranslator::DxbcShaderTranslator(uint32_t vendor_id,
-    : edram_rov_used_(edram_rov_used) {
+                                           bool edram_rov_used)
+    : vendor_id_(vendor_id), edram_rov_used_(edram_rov_used) {
   // Don't allocate again and again for the first shader.
   shader_code_.reserve(8192);
   shader_object_.reserve(16384);
@@ -519,6 +522,11 @@ void DxbcShaderTranslator::Reset() {
   std::memset(&stat_, 0, sizeof(stat_));
 }
 
+bool DxbcShaderTranslator::UseSwitchForControlFlow() const {
+  // Xenia crashes on Intel HD Graphics 4000 with switch.
+  return FLAGS_dxbc_switch && vendor_id_ != 0x8086;
+}
+
 uint32_t DxbcShaderTranslator::PushSystemTemp(bool zero) {
   uint32_t register_index = system_temp_count_current_;
   if (!IndexableGPRsUsed() && !is_depth_only_pixel_shader_) {
@@ -1208,7 +1216,7 @@ void DxbcShaderTranslator::StartTranslation() {
   ++stat_.instruction_count;
   ++stat_.dynamic_flow_control_count;
   // Switch and the first label (pc == 0).
-  if (FLAGS_dxbc_switch) {
+  if (UseSwitchForControlFlow()) {
     shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_SWITCH) |
                            ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(3));
     shader_code_.push_back(
@@ -6427,7 +6435,7 @@ void DxbcShaderTranslator::CompleteShaderCode() {
     // closing upper-level flow control blocks.
     CloseExecConditionals();
     // Close the last label and the switch.
-    if (FLAGS_dxbc_switch) {
+    if (UseSwitchForControlFlow()) {
       shader_code_.push_back(
           ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_BREAK) |
           ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(1));
@@ -7982,7 +7990,7 @@ void DxbcShaderTranslator::ProcessLabel(uint32_t cf_index) {
   // execs across labels.
   CloseExecConditionals();
 
-  if (FLAGS_dxbc_switch) {
+  if (UseSwitchForControlFlow()) {
     // Fallthrough to the label from the previous one on the next iteration if
     // no `continue` was done. Can't simply fallthrough because in DXBC, a
     // non-empty switch case must end with a break.
@@ -8067,7 +8075,7 @@ void DxbcShaderTranslator::ProcessExecInstructionEnd(
   if (instr.is_end) {
     // Break out of the main loop.
     CloseInstructionPredication();
-    if (FLAGS_dxbc_switch) {
+    if (UseSwitchForControlFlow()) {
       // Write an invalid value to pc.
       shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
                              ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(5));

src/xenia/gpu/dxbc_shader_translator.h

@@ -24,7 +24,7 @@ namespace gpu {
 // Generates shader model 5_1 byte code (for Direct3D 12).
 class DxbcShaderTranslator : public ShaderTranslator {
  public:
-  DxbcShaderTranslator(bool edram_rov_used);
+  DxbcShaderTranslator(uint32_t vendor_id, bool edram_rov_used);
   ~DxbcShaderTranslator() override;
 
   // Constant buffer bindings in space 0.
@@ -758,6 +758,9 @@ class DxbcShaderTranslator : public ShaderTranslator {
     return is_depth_only_pixel_shader_ || is_pixel_shader();
   }
 
+  // Whether to use switch-case rather than if (pc >= label) for control flow.
+  bool UseSwitchForControlFlow() const;
+
   // Allocates a new r# register for internal use and returns its index.
   uint32_t PushSystemTemp(bool zero = false);
   // Frees the last allocated internal r# registers for later reuse.
@@ -994,6 +997,9 @@ class DxbcShaderTranslator : public ShaderTranslator {
   // Buffer for instruction disassembly comments.
   StringBuffer instruction_disassembly_buffer_;
 
+  // Vendor ID of the GPU manufacturer, for toggling unsupported features.
+  uint32_t vendor_id_;
+
   // Whether the output merger should be emulated in pixel shaders.
   bool edram_rov_used_;
 

src/xenia/gpu/shader_compiler_main.cc

@@ -103,7 +103,7 @@ int shader_compiler_main(const std::vector<std::wstring>& args) {
         GlslShaderTranslator::Dialect::kGL45);
   } else if (FLAGS_shader_output_type == "dxbc") {
     translator =
-        std::make_unique<DxbcShaderTranslator>(FLAGS_shader_output_dxbc_rov);
+        std::make_unique<DxbcShaderTranslator>(0, FLAGS_shader_output_dxbc_rov);
   } else {
     translator = std::make_unique<UcodeShaderTranslator>();
   }

src/xenia/ui/d3d12/d3d12_provider.cc

@@ -11,6 +11,9 @@
 
 #include <gflags/gflags.h>
 
+#include <malloc.h>
+#include <cstdlib>
+
 #include "xenia/base/logging.h"
 #include "xenia/ui/d3d12/d3d12_context.h"
 
@@ -125,8 +128,7 @@ D3D12Provider::InitializationResult D3D12Provider::Initialize() {
     return InitializationResult::kDeviceInitializationFailed;
   }
 
-  // Choose the adapter and create a device with required features.
+  // Choose the adapter.
-  // TODO(Triang3l): Log adapter info (contains a wide string).
   uint32_t adapter_index = 0;
   IDXGIAdapter1* adapter = nullptr;
   while (dxgi_factory->EnumAdapters1(adapter_index, &adapter) == S_OK) {
@@ -158,6 +160,24 @@ D3D12Provider::InitializationResult D3D12Provider::Initialize() {
     dxgi_factory->Release();
     return InitializationResult::kDeviceInitializationFailed;
   }
+  DXGI_ADAPTER_DESC adapter_desc;
+  if (FAILED(adapter->GetDesc(&adapter_desc))) {
+    XELOGE("Failed to get the DXGI adapter description.");
+    adapter->Release();
+    dxgi_factory->Release();
+    return InitializationResult::kDeviceInitializationFailed;
+  }
+  adapter_vendor_id_ = adapter_desc.VendorId;
+  size_t adapter_name_length =
+      std::wcstombs(nullptr, adapter_desc.Description, 0);
+  char* adapter_name_mbcs =
+      reinterpret_cast<char*>(alloca((adapter_name_length + 1) * sizeof(char)));
+  std::wcstombs(adapter_name_mbcs, adapter_desc.Description,
+                adapter_name_length + 1);
+  XELOGD3D("DXGI adapter: %s (vendor %.4X, device %.4X)", adapter_name_mbcs,
+           adapter_desc.VendorId, adapter_desc.DeviceId);
+
+  // Create the Direct3D 12 device.
   ID3D12Device* device;
   if (FAILED(pfn_d3d12_create_device_(adapter, D3D_FEATURE_LEVEL_11_0,
                                       IID_PPV_ARGS(&device)))) {

src/xenia/ui/d3d12/d3d12_provider.h

@@ -62,6 +62,9 @@ class D3D12Provider : public GraphicsProvider {
     return start;
   }
 
+  // Adapter info.
+  uint32_t GetAdapterVendorID() const { return adapter_vendor_id_; }
+
   // Device features.
   uint32_t GetProgrammableSamplePositionsTier() const {
     return programmable_sample_positions_tier_;
@@ -126,6 +129,8 @@ class D3D12Provider : public GraphicsProvider {
   uint32_t descriptor_size_rtv_;
   uint32_t descriptor_size_dsv_;
 
+  uint32_t adapter_vendor_id_;
+
   uint32_t programmable_sample_positions_tier_;
   bool rasterizer_ordered_views_supported_;
   uint32_t tiled_resources_tier_;